1 

 

PharmaSUG2011 – Paper HS03

Using SAS Predictive Modeling to Investigate

the Asthma’s Patient Future Hospitalization Risk Yehia H. Khalil, University of Louisville, Louisville, KY, US

ABSTRACT The focus of this study is to develop predictive models to forecast future Asthma hospitalizations for patients diagnosed with Asthma. Using SAS® Software tools such as predictive models provides great benefit to insurance companies, hospitals, and pharmaceutical companies trying to manage their resources and define future plans and policies. The strength areas of predictive modeling are: a) an ability to incorporate any type of variable into the analysis, b) Dynamic, as they can easily accommodate any information as they become available to adjust the model accordingly. A recent study shows that around 21 million Americans are diagnosed with asthma, which represents a large section of the insurance and pharmaceutical patient population. Asthma patients are classified into two groups: Controlled vs. Uncontrolled asthma; however, during the year 2000, the national surveys show that there were nine million physician office visits and more than 900,000 hospitalization cases in addition to the emergency room visits. Based on the American Lung Association reports, in 2007, almost 2.5 million people over the age of 65 had asthma, and more than 1 million had an asthma attack or episode.( American Lung Association 2007) We examine data from Medical Expenditure Panel Survey (MPES) and California Health Interview Survey. It was proven that the use of predictive modeling did show benefit by pointing to optimal treatment for patients and can be used to develop better investments for health plans.

INTRODUCTION Asthma is a chronic inflammatory disorder of the airways, designated as ICD9-CM-493 in the (ICD-9 classification system. Asthma is a chronic disease that cannot be cured. Asthma is ranked as one the most common diseases within the United States. It is usually treated by two types of medications: long term and quick relief medication. Missing work, missed school days are examples for the indirect cost while treatments and hospitalization represent the direct cost. Asthma patients use a significant portion of insurance companies, state insurance funds and hospital resources. Asthma hospitalization rates are growing, especially for children. Since Asthma patient are classified as high cost patients, as the Urban Institute’s health costs survey estimates that an asthma patient can spend as much as $4,900 many research studies were directed at hospitalizations to decrease the healthcare bill. (Health Costs Survey 2005) Medications are the prime cost components for the asthma patient; for example, for inpatients, asthma represents 30% of the overall cost. Asthma guidelines generally use an Expert panel, and have been developing over the years to the current state as consensus panel guidelines. Since guidelines rely largely on observational studies, the definitions are shaped by consensus; while it seems reasonable that there should be a relationship between observations and consensus, this does not necessarily occur. On the other hand, predictive modeling provides a great opportunity for better healthcare planning, management and clinical practice guidelines. Identifying the patients who will need medical attention in the future enhances health care management guidelines, which exist to define standard treatments for patient and to reduce the utilization of resources. Generally speaking, asthma affects children and African Americans within the general population and affects females more than males within the adult population. Data collected over the last few years show that asthma is ranked as one of the top illnesses that affect the American people. Almost 15 million people are diagnosed with asthma, The National Heart, Lung and Blood Institute states that costs for Asthma are approximately $14.0 billion with $9.4billion in direct costs, $2.7billion for morbidity, and $1.9billion for mortality. Adults or children who are diagnosed with asthma miss work, school days, and also sleep poorly at night, and in many cases, asthma can affect their mobility. On the other hand, several things can affect the patient's general health condition such as nutrition, smoking, and exercise. Uncontrolled asthma can lead to several complications such as school absenteeism, breathing difficulty, hospitalization and death, in some cases. The main focus of service providers, payers and decision-makers is to minimize the number of emergency room visits and the high cost of hospitalizations. Several intervention approaches are used to ensure that goal. However,

2 

 

the large number of patients highlights the need for efficient resource management and the need to predict the future health requirements to avoid a resources shortage, as the main challenge is to categorize patients who will need treatment. To achieve this goal, decision-makers want to investigate several issues, such as the number of patients with a controlled condition vs. uncontrolled, geographical patient distribution, interventions and so on. Predictive modeling is a powerful tool to enhance the decision making process; in the healthcare domain, it is a relatively new concept. Predictive modeling calculates the probability of future adverse events for any patient based on past clinical information compared to the population. The model we define is created using several factors, or “predictors,” which are expected to affect the future of this patient status, such as age, gender, and medical history. Predictive modeling tools are used for several applications such as:

1- Predicting future costs and the design of benefits packages a. deductibles b. co-pays

2- Patient profiling a. Identification of high-risk patients b. Disease management and case management c. cost savings

3- Provider profiling/Plan profiling 4- Risk adjustment

For the last few years, predictive modeling accuracy has increased and proven to improve healthcare management. It can be used to develop better investments for health plans and enhance resources utilization. The strength areas of predictive modeling are:

a) The ability to incorporate any type of variable into the analysis b) Dynamic, as a model can easily accommodate any information as it becomes available to adjust the model

accordingly.

USING PREDICTIVE MODELING Predictive Modeling is one of the two main branches of data modeling; it is a process to create a statistical model of future behavior. Regression and Classification demonstrate the main idea of predictive modeling (Cerrito 2006)  

‐ If the Response variable is numerical, predictive modeling is called regression.

‐ If the Response variable is nominal, predictive modeling is called Classification. Developing predictive models includes data preparation, prediction, and the analysis of results. The model tuning process is a time consuming processes. The main challenges that the data analyst encounters are

‐ Selecting the predictive method

‐ Selecting the independent variables, or predictors Regression models are the backbone of predictive modeling, Regression is the process of establishing a mathematical model as a function to represent the relation between the different variables (predictors). There is a wide-range of models that can be applied while performing predictive analytics. Some of them are:

‐ Linear regression model

‐ Logistic regression

‐ Multinomial logistic regression Artificial Neural networks (ANN) provide an information processing model for computing that involves developing mathematical structures with learning. ANNs are used to provide projections given new situations of interest and to answer "what if" question. The following figure shows the ANN diagram.

3 

 

Decision Tree is a predictive modeling tool that does prediction by classification approach. The key element for building a decision tree is finding the best rule that can spilt the data record in the best manner. At each level, the best splitting rule is applied until you cannot find better ways to split the input records. SAS® SEMMA METHODOLOGY The SEMMA methodology was introduced by SAS® technology; SEMMA represents the lifecycle of the core process of data mining. It also represents the complete functionally of SAS Enterprise Miner, and is acronym for sample, explore, modify, model, and assess; SEMMA can be described briefly as follows:(Sumathi, 2006)

Sample: Extracting a portion of a large data set big enough to contain the significant information, but still manageable. In addition, it supports a data partition into Training, Validation, and Testing subsets.

Explore: To enable you to become more familiar with your data and to detect any pattern or any anomalies. This process can be done visually or using techniques such as clustering, path analysis and text analysis.

Modify: Data may need to be modified for a number of reasons; since data mining is a dynamic process, then the objective can change and the data will be modified. The other reason to modify data is to utilize data discoveries.

Model: it the process of using current data to investigate the relationship between the parameters to create a model capable of future behavior. Several techniques can be used such as neural network, tree based and regression models.

Assess: it is very important to be able to test the usefulness and reliability of a model, which can be done easily by testing model against known data.

DATA SUMMARIZATION 

Input layer

Hidden layer

Output layer

Sample

Explore

Modify

Model

Access

Figure 1. Artificial Neural networks

Figure 2. SAS® SEMMA METHODOLOGY

4 

 

To illustrate the asthma hospitalizations risk problem, the Medical Expenditure Panel Survey (MEPS) data sets were used to find general conclusions about asthma in the United States and the California Health Interview Survey (CHIS) data sets were used to investigate predictive modeling. As shown by the following table (Table 1) 13.6% of the population has been diagnosed with asthma at least once based on the CHIS 2009 data sets. (http://www.chis.ucla.edu/) For this study we focused on the CHIPS dataset for the year 2009. It surveyed 47,614 adults, 3,379 adolescents and 8,945 children. The survey includes robust samples of various races. We tested several subsets of parameters or factors to be tested versus the hospitalization risk or need for emergency room visits or the frequency of attacks. In this study, we included many parameters as follows:

Demographic Information: Age, Gender, Race, Marital Status. Health Behaviors: physical activities, Fast food, Alcohol consumption. Health Conditions other than Asthma. Health Insurance. Poverty level. Emergency Preparedness Module: Medication Mental or emotional condition.

Table1. Patient diagnosed with asthma summary

Male Female All

Ever diagnosed with asthma No. of Obs. % No. of Obs. % No. of Obs. %

Has asthma 2,401,000 13.4 2,543,000 13.9 4,944,000 13.6

Does not have asthma 15,574,000 86.6 15,747,000 86.1 31,321,000 86.4

TOTAL 17,975,000 100 18,290,000 100 36,265,000 100 The following figure (Figure 1) shows that the number of females diagnosed with asthma is higher than for males. Also, the number of females who no longer have asthma is smaller compared to the number for males.

 Figure 3. Patient diagnosed with asthma at least once  Another aspect to check is marital status, based on the CHIS classification, which uses four main categories as shown in the following table (Table 2). Table 2: Marital status summary

Male Female All

Marital status No. of Obs. % No. of Obs. % No. of Obs. %

Married 7,461,000 56.6 7,316,000 53.4 14,777,000 55

Live with partner 1,002,000 7.6 976,000 7.1 1,978,000 7.4

Separated / Divorced / Widowed 1,225,000 9.3 2,641,000 19.3 3,866,000 14.4

Single, never married 3,490,000 26.5 2,763,000 20.2 6,253,000 23.3

TOTAL 13,179,000 100 13,695,000 100 26,874,000 100

5 

 

The insurance coverage is a very important element for the patient to maintain an acceptable level of medical treatment, in particular, for out-patients. The following table (Table 3) represents summary for patients who had no insurance at all for the last 12 months, either partial or complete coverage. Table 3: Medical insurance coverage summary

During the last 12 months

No insuranceat all

Insurance fora part

Had insuranceentire time

Currently have asthma % % % Currently has asthma 48.8 56.7 60.6 No longer has asthma 51.2 43.3 39.4

 It is known that the asthma patient can experience work or other activities limitations; the following table (Table 4) illustrates the scope of the problem. For example, 58.8% of the patients who had 1-2 days of work limitation had it caused by asthma. Table 4: Work/other activates limitation caused by asthma

During the last month None 1 - 2 days 3 - 5 days 6 - 12 days 13 - 29 days 30 days Caused by asthma 59.1% 58.8% 61.3% 65.9% 64.3% 67.1%

Data show that children and senior patients are more affected by asthma as shown in the following table (Table 5). Table 5: Age categories

During the last month

Child (0-11)

Adolescent (12-17)

Adult(18-64)

Senior (65+)

Diagnosed asthma 65.4% 57.8% 58.2% 66.4% USING SAS® PREDICTIVE MODELING The following figure (Figure 4) shows the data diagram for predictive modeling using SAS® Enterprise Miner. As a data preprocessing step, data were filtered to include only patients diagnosed with Asthma using the ICD-9 code of 493. Data were partitioned using the data partition node with percentages 40% training, 30% testing, and 30% validation. (Sarma, 2006) 

Figure 4. Predictive Modeling Diagram Note: use version 5 or 6, not version 4.3 for this diagram One of the major problems of using predictive modeling is that for some applications, certain models will not work, yet SAS software provides an easy way to use several models such as neural network, regression and decision. In addition, it provides an assessment node to compare each model result and to chose the best one. For the first round, we used all of the 473 response parameters as input and the occurrence of hospitalization as the outcome variable. It showed that several factors were related:

ER visits within the last year Asthma episode/attacks in the past 12 months How often had asthma symptoms Others

It is valid that those parameters have influence on the occurrence of the hospitalization event; however, this study’s objective is to go beyond that and to investigate other factors affecting the hospitalization event. Within the decision

6 

 

tree model, a competing splitting feature provides an easy way to explore other splitting factors. For example, the following table (Table 6) shows the competing splitting vs. the ER visits with the past year. Table 6: Competing splitting

Variable Logworth Groups LabelER 24.694 2 ER VISIT WITHIN THE PAST YEAR AB1 13.429 2 GENERAL HEALTH CONDITION AB42 12.731 2 WORKDAYS MISSED DUE TO ASTHMA IN PAS AD52 10.549 2 HAS DIFFICULTY DRESSING, BATHING, GETTI AB41   9.134   2 ASTHMA EPISODE/ATTACK IN PAST 12 MOS

Logworth is a split decision statistic that measures the effectiveness of a particular split decision at differentiating values of the target variable. For the next round of the run, the “expected” factors were removed to give more opportunity for the competing splitting. The second round showed parameters such as

Taking daily medication to control asthma General health conditions Type 1 or Type 2 diabetes   Psychological Distress   Poverty level  

To achieve a better understanding of the problem, we used a backtracking technique. For instance, tracking daily medication to control asthma is very significant to avoid future hospitalizations. Rx-Coverage, disability, language barriers and some mental problem can prevent asthma patients from taking the medication regularly. The following figures show sample output and results for predictive modeling models: Decision tree, Regression and neural networks.

Figure 5. Sample output of decision tree model  

 Figure 6. Fit statistics for regression model  

7 

 

Figure 7 Fit statistics for neural network model The above figures (Figure 6 & Figure 7) show fit statstics for the neural network model and the regression model, selected statstics can be used as the modeling selection criteria statistics such as: average profit and misclassification rate. While the average error is the modeling selection criteria statstics for chosing the best neural network mode. If the modeling t statistics are totlay changed between the partitioned data sets then it will be an indication of overfitting to the regression model. Another important feature of a SAS® Enterprise Miner solution offers is the assessment node, where it compares ASE (Average Square Errors), Schwarz Bayesian criterion and the misclassification rate. In many cases, the neural network model will have a slightly smaller ASE and misclassification rate. The main concern with neural network is that they are hard to explained. CONCLUSION   SAS® Enterprise Miner provides a powerful environment for predicative modeling by supporting several models and the ability to compare between them. The use of predicative modeling for solving healthcare problems is not new, yet this study illustrates a new application, risk of future hospitalization for asthma patients. The results show that predictive modeling can enhance the decision making process and future planning for hospitals’ resource utilization. Also, it was clear that by using predictive modeling results in terms of which parameters influence future hospitalizations, it is possible to enhance interventions, programs and alternatives to avoid future hospitalizations. Results show that the general health conditions, psychological distress and poverty level affect future hospitalization risk. In addition, the Rx coverage and patient's disability influence taking medication regularly and can increase the future hospitalization risk. REFERENCES 1. American Lung Association. Epidemiology & Statistics Unit, Research and Program Services. Trends in Asthma

Morbidity and Mortality, November 2007. 2. Health Costs Survey 2005. 3. Patricia B. Cerrito: Introduction to data mining using SAS Enterprise Miner, SAS Institute, 2006. 4. S. Sumathi, S. N. Sivanandam: :Introduction to data mining and its applications, Springer 2006. 5. Kattamuri S. Sarma: Predictive Modeling With SAS Enterprise Miner: Practical Solutions for Business

Applications, SAS Institute , 2006. 6. National Asthma Education and Prevention Program Expert Panel Report 3: Guidelines for the Diagnosis and

Management of Asthma. Rockville, MD. National Heart, Lung, and Blood Institute, US Department of Health and Human Services; 2007. NIH publication 08-4051.

7. Woodfield, Terry: "Data Conversion Pitfalls", Proceedings of Computer Science and Statistics: 26th Symposium on the Interface, 1994, pp. 362-371.

8. MC Weinstein, EL Toy, EA Sandberg, PJ Neumann, JS: Modeling for Health Care and Other Policy Decisions: Uses, Roles, and Validity, Value in Health, Blackwell Synergy, 2001.

ACKNOWLEDGMENT The author would like to thanks Prof. Patricia Cerrito of the University of Louisville for her helpful suggestions and advices. CONTACT INFORMATION Your comments and questions are valued and encouraged. Contact the author at:

Yehia H. Khalil University of Louisville, Louisville KY Email: Yehia.khalil@ieee.org